Wheels for linux aarch64

[thomasaarholt]

Hello! If this is still in development, I'd love to see wheels for linux aarch64, so that this can be installed without rust in Docker on my M1 mac :)

[kylebarron]

👋 I haven't had much free time the last couple months, but I want to get back to this. (I made a polars PR for fixed size list types, and I want to push that through, which should unblock a couple things here.)

In terms of the linux aarch wheels; do you have an example for that? I'm guessing it has to be built with QEMU? Are there manylinux-specific docker containers for aarch? It looks like there are some here: https://github.com/pypa/manylinux. So it might be as simple as copying this block

geopolars/.github/workflows/python-wheels.yml

Lines 13 to 19 in 4ea84f4

	build-wheel-manylinux-2-24:
	name: "Build manylinux 2.24 wheels"
	runs-on: ubuntu-latest
	container: quay.io/pypa/manylinux_2_24_x86_64
	defaults:
	run:
	working-directory: py-geopolars

to include an aarch docker container as well

[thomasaarholt]

Yes, I think that's a good starting point.

For how polars does it: https://github.com/pola-rs/polars/blob/5c61983dba25724ae0252eec3b46b877c0151a57/.github/workflows/create-python-release.yml#L38

[thomasaarholt]

Very excited to hear that this project is still in development 😊 I've worked on a data-science project and have had to resort to resort to converting my geography-columns back and forth between LineStrings and List[List[Float64]] in order to pass things between pandas and polars.

[kylebarron]

For how polars does it: pola-rs/polars@5c61983/.github/workflows/create-python-release.yml#L38

That's a good reference; I don't currently use maturin-action in the geopolars actions because I'm also compiling proj from source, and it's not clear maturin-action has a way to do that

[kylebarron]

geography-columns back and forth between LineStrings and List[List[Float64]]

The List[Float64] isn't going to be that efficient because you're storing an extra offsets column for every coordinate, so hopefully fixed size lists will make that better.

Are you using rust or Python? I.e. are you converting between shapely linestrings or georust linestrings?

[thomasaarholt]

you're storing an extra offsets column for every coordinate

That's interesting, not something I had considered!

I'm doing this in python, converting shapely linestrings. My data is thankfully not that large, at a few thousand nodes, so conversion isn't a slow process.

[kylebarron]

At that point it might be better to use a wkb column of data when you want it in the polars dataframe and convert it to shapely geometries when you want to do operations. The WKB is probably more space efficient than the non-fixed size list and the shapely to/from wkb is more optimized I think